Research And Application Of Intelligent Dehazing Model For Optical Remote Sensing Images Considering Atmospheric Light Intensity

With the development of space remote sensing and information technology,remote sensing imagery has entered a new stage of comprehensive application,and is widely used in such fields as geographic data acquisition,earth resources acquisition,military,disaster emergency monitoring and land change detection.However,in cloudy conditions,loss of details in the images of the observed scenes is usually due to the attenuation of light energy on the imaging channel,which is visually manifested as poor visibility and reduced contrast.These cloud images not only reduce the quality of remote sensing images,but also affect the accuracy of subsequent remote sensing application tasks such as semantic segmentation,object detection,change detection.Remote sensing images are inherently massive,multi-dimensional and time-sensitive.In recent years,during while deep learning algorithms were emerging in the field of artificial intelligence,they also had an advantage in processing remote sensing data,learning high-level abstract features from massive training data and greatly improving the accuracy of predictions.Therefore,in conducting intelligent optical remote sensing image de-hazing models research is of great practical significance in the improvement of efficient remote sensing image quality.In this thesis,several theories and algorithms for the de-hazing of single-temporal remote sensing images and information compensation in the cloudy areas of multi-temporal remote sensing images have been studied in depth and corresponding de-hazing models have been designed for the de-hazing of optical remote sensing images.As such taken into account was the influence of clouds and haze on the imaging process of those images,on such as,transmittance map estimation based on atmospheric scattering model,deep fully convolution network model.The main work in the thesis,includes the following aspects.(1)A Summary of the physical imaging principle of clouds and haze in remote sensing images,introduction of the widely used atmospheric scattering model in de-hazing models,and the proposition of the concept of transmittance maps,the simplification of the problem of solving de-hazing images to the problem of solving transmittance map;the introduction of several more innovative de-hazing algorithms,based on atmospheric scattering models,and an analysis of their concepts and shortcomings,to provide necessary theoretical support for subsequent model research.(2)The Combined Regression De-haze Network(CRDN)model of atmospheric light intensity and transmittance maps based on the fully convolutional network model was proposed and designed to solve the single-temporal image de-hazing problem.This model obtains haze-free images by estimating both transmittance maps and atmospheric light intensity.Experiments have shown that the model produces a de-hazing image with a tone close to that of a haze-free image,and hence can well avoid under-dehazing or "over-dehazing".(3)To evaluate the applicability of the de-hazing model,the following are introduced: multi-scale structural similarity(MS-SSIM),feature similarity(FSIM)and quantitative gradient similarity(GSM)to further optimize the above model.The experiments show that the optimized models are both suitable for the actual requirements of the de-hazing task,and also facilitation of a more objective and comprehensive evaluation of the inter-model merits of different de-hazing models.To further validate the effectiveness of the CRDN model,it was a)applied to the preprocessing module of the semantic segmentation model,b)it experimentally demonstrated that a slight cloud effect can lead to a significant reduction in the accuracy of semantic segmentation models,and c)it revealed that the remote sensing images dehazed by the CRDN model could obtain better accuracy in further remote sensing image interpretation tasks.(4)An intelligent CRDN+ model structure is proposed and designed for semantic segmentation,to improve,in practice,the application efficiency of the model,on the basis of the extensive application of the pre-processing of satellite remote sensing images and UAV images in harsh conditions,and a unique set of loss functions is designed for this model.The new model can be used for preprocessing modules of remote sensing image interpretation tasks such as semantic segmentation.When no additional fine-tune training is conducted,the module effect is equivalent to the effect of semantic segmentation after de-hazing using the CRDN model;however,after fine-tune training,the accuracy of semantic segmentation can be significantly improved.(5)To remove thick haze from multi-temporal images,a cloud area information compensation model based on a transmittance map is proposed and designed.This model combines the cloud detection task with the under-cloud pixel recovery task,quantifies the influence of cloud and haze on optical remote sensing images as a weight function,and restores the cloud image by combining all the image information of each phase,in the same area,Thus the problem of cloud area information compensation for optical remote sensing images well is solved.